Jerusalem 1840-1949 Road Extraction and Alignment

Motivation

The creation of large digital databases on urban development is a strategic challenge, which could lead to new discoveries in urban planning, environmental sciences, sociology, economics, and in a considerable number of scientific and social fields. Digital geohistorical data can also be used and valued by cultural institutions. These historical data could also be studied to better understand and optimize the construction of new infrastructures in cities nowadays, and provide humanities scientists with accurate variables that are essential to simulate and analyze urban ecosystems. Now there are many geographic information system platforms that can be directly applied, such as QGIS, ARCGIS, etc. How to digitize and standardize geo-historical data has become the focus of research. We hope to propose a model that can associate geographic historical data with today's digital maps, analyze and study them under the same geographic information platform, same coordinate projection, and the same scale. Eventually, it can eliminate errors caused by scaling, rotation, and the deformation of the map carrier that may exist in historical data and the entire process is automated and efficient.

The scale is restricted to Jerusalem in our project. Jerusalem is one of the oldest cities in the world and is considered holy to the three major Abrahamic religions—Judaism, Christianity, and Islam. We did georeferencing among Jerusalem’s historical maps from 1840 to 1949 and the modern map from OpenStreetMap so that the overlaid maps will reveal changes over time and enable map analysis and discovery. We focused on the wall of the Old City as the morphological feature to do georeferencing because the region outside the Old City has seen many new constructions while the Old City has not great changes and the closed polygon of the wall is relatively more consistent than other features like road networks. More specifically, we used dhSegment, a historical document segmentation tool, to extract the wall of the Old City of Jerusalem and proposed an alignment algorithm exploiting the geometrical features of the wall to align the maps.

Deliverables

- The latitudes and longitudes of the four vertices of the raw maps.

- Preprocessed (cropped and resized) patches of the raw maps with images information (including the position of the sub-image in the raw map) stored in a csv file.

- 42 1000 * 1000 annotations of the Old City area.

- Trained segmentation model.

- Code and instructions to reproduce the results on github.

Methodology

Dataset

• 126 historical maps of Jerusalem from 1837 to 1938.

• Modern geographical data of Jerusalem from OpenStreetMap.

Wall Extraction

We first used dhSegment to do image segmentation in order to extract the wall polygon from the maps. dhSegment [1] is a generic approach for Historical Document Processing, which relies on a Convolutional Neural Network to predict pixelwise characteristics. The whole process include preprocessing, training, predicting and postprocessing.

Figure 2: network architecture

• Preprocessing

To make the image data fit into the neural networks and to make the old city region more dominant in the image, we cropped and scaled the original image by hand for subsequent uses.

• Map annotating

We used Procreate to pixel-wisely label 42 maps according to whether the pixel belong to the old city with different colors (RGB(0, 0, 0) for pixels inside the wall, RGB(255, 0, 255) for pixels outside the wall).

• Training

We divided the 42 annotated patches into 36 and 6 for training and validating respectively, and trained the model with learning rate 5e-5, batch size: 2, N_epochs: 40, and data augmentation like rotation, scaling and color.

Figure 3: Original Image

Figure 4: Labelled Image

• Prediction

By doing prediction on the testing set, we got the predicted old city region and naturally got the contour of it, i.e. the wall.

Figure 5: Original Image

Figure 6: Predicted Binarized Mask

• Postprocessing

We removed the minor noises by retain the largest connected region and drew the contour of it. We put the contour on the original image and find that it generally fits well, although there are some minor deviations.

Figure 7: After removing noises and drawing contour

Wall Alignment

After obtaining the wall polygons, we try to fit them to the the reference wall polygon from the OpenStreetMap and obtain the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.

• Overview

The overall algorithm proposed is shown as follows. We do scaling, transformation and rotation iteratively for a given iteration numbers in order to get a better alignment result since the initial scaling isn't the accurate enough with the angle difference of the source polygon and reference polygon.

For i = 0 to iters :

​Estimate and do scaling

​​::For each point pair P in selected point pair set:

Calculate and do transformation

For k = -6 to 6 :

Do rotation of angle = k(iters - i) around P

Calculate overlapping area

Update maximum overlapping area

• Scaling

The method to do scaling is shown in the image below.

Figure 8: An approximate scaling

• Translation & Rotation

The goal of the translation and rotation is to obtain optimal translation and rotation to maximize the overlapping area of the two polygons. Since it’s computationally expensive to search for the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space to translations that can fit the key point pairs respectively together (up-left points, up-right points, bottom-left points, bottom-right points and the centroids) and rotation angle set, which starts with [-30, 30] of step 5 and whose range and step will get smaller after every iteration (it's intuitive since after one iteration, we will need more accurate rotation angles).

For every translation, we fix the key point together, rotate the wall and calculate the overlapping area. To calculate the overlapping area, we use the FloodFill algorithm to fill the area bounded by the source polygon and reference polygon respectively and use the logical-and operation to obtain the overlapping region. We then count the number of pixels that are not zero to be the overlapping area.

After the iterations, we get the maximized overlapping area and the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.

Figure 9: Translation and rotation

The translation and rotation step is to find the best translation and rotation angle that maximizes the overlapping area of the two walls. Since it’s computationally expensive to search the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space. We limit the search space of translation by only considering the translations that can fit the key point pairs, (P_{00}, P_{01}), (P_{10}, P_{11}), (P_{20}, P_{21}), (P_{30}, P_{31}), together. Here we can also use more key points, like the centroids. For every translation, we fix the key point and rotate the wall every 5 degrees from -30 degrees to 30 degrees and calculate the overlapping area. To calculate the overlapping area, we use the FloodFill algorithm to fill the area bounded by the walls respectively and use the logical-and operation to obtain the overlapped region and then see the number of pixels that are not zero to be the area. After the iterations, we get the maximized overlapping area and the transformation matrix.

Coordinate System Transformation

Since the final goal is to get the latitudes and longitudes in the geographic coordinate system of the four vertices of the raw maps, we finally transformed the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps through mathematical computations.

Figure 10: Coordinate System Transformation

Results

Wall Extraction

For the wall extraction step, the neural network generally gives good predictions.

Figure 11: 1837 contour

Figure 12: 1845 contour

Figure 13: 1883 contour

Figure 14: 1921 contour

But for some maps, there are some concavity and convexity in the predictions, but these flaws won't influence the alignment as shown in the next section.

Figure 15: 1853 contour

Figure 16: 1855 contour

Figure 17: 1920 contour

Figure 18: 1929 contour

Wall Alignment

For good predictions in the previous segmentation step, the alignment algorithm can give a satisfying alignment.

Figure 19: 1883 contour	Figure 20: 1921 contour
Figure 21: 1883 predict	Figure 22: 1921 predict

The bias generally is caused by the intrinsic difference of the wall in the original map and the reference wall. The maps whose wall polygons have large differences from the nowadays wall polygon were generally produced in the early 19 century. Take the map in 1837 and 1845 for example.

Figure 23: 1837 contour	Figure 24: 1845 contour
Figure 25: 1837 predict	Figure 26: 1845 predict

We add the main road from OpenStreetMap to the reference map and align them to the historical maps. The result also shows that the alignment results of more recent maps are better.

Figure 27: 1845 alignment results

Figure 28: 1887 alignment results

Figure 29: 1915 alignment results

We plotted the overlapping area ratio with regard to the year of the map and find that the alignment result of early 19 century's maps are not as good as that of later maps, which is consistent with our previous observations.

Figure 30: Overlapped area ratio w.r.t the year of map

Even for some maps with not that satisfying prediction results, the robustness of the alignment algorithms proposed make the polygons generally align together, thus the coordinates obtained are still reliable.

Figure 31: 1853 contour	Figure 32: 1855 contour	Figure 33: 1920 contour	Figure 34: 1929 contour
Figure 35: 1853 predict	Figure 36: 1855 predict	Figure 37: 1920 predict	Figure 38: 1929 predict

Coordinate System Transformation

The results are stored in a csv file shown as follows.

Figure 39: Coordinate transformation

Limitations

- Although the alignment algorithm is not sensitive to the minor flaws of the segmentation results, the alignment performance is still subject to the segmentation results. Specifically, when the segmentation gives very bad predictions, the alignment algorithm just can't work.

- The project can only do linear transformation, i.e. scaling, translation and rotation, but it can't do deformation.

- Only the closed polygon of the wall are utilized and segmented. The morphological feature of road networks and buildings can also be exploited in future works.

Project Plan and Milestones

Milestones

• Do survey on semantic segmentation and registration algorithms, get familiar with our dataset and decide what feature to be used for alignment.

• Preprocess the dataset.

• Annotate the maps and use dhSegment for segmentation.

• Design and implement the alignment algorithms.

• Do coordinate system transformation.

• Further improve the robustness of the alignment algorithm and connect the processing steps.

• Sort out the code and write the report.

Plans

Date	Task	Completion
By Week 4	Brainstorm project ideas. Prepare slides for initial project idea presentation.	✓
By Week 6	Study relevant works about road extraction and alignment. Use Procreate to produce road-annotated images as the training dataset. Study and apply Projet JADIS to our dataset.	✓
By Week 8	Filter and preprocess the dataset for training and testing. Try various key point selection algorithms like SIFT and SURF. Try various existing point cloud registration algorithms like ICP and TEASER-plusplus. Determine to use the wall for alignment. Design and develop wall alignment algorithms.	✓
By Week 10	Use Procreate to produce wall-annotated images as the training dataset. Train the dhSegment model from scratch and use it for wall extraction. Implement the image postprocessing methods. Further improve the performance of the alignment algorithm. Prepare for the midterm presentation.	✓
By Week 11	Produce more wall-annotated images. Fine-tune the dhSegment model to get better segmentation results.	✓
By Week 12	Further improve the performance and robustness of the alignment algorithm to make it robust to poorly predicted data. Transform the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps. Make all the processing steps a pipeline and user-friendly.	✓
By Week 13	Sort out the codes and push them to GitHub repository. Write the project report. Prepare slides for final presentation.	✓
By Week 14	Finish presentation slides and report writing. Presentation rehearsal and final presentation.	--

Github Link

https://github.com/mzp0514/JerusalemGeoreferencing

References

[1] Oliveira, Sofia Ares, Benoit Seguin, and Frederic Kaplan. "dhSegment: A generic deep-learning approach for document segmentation." 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR). IEEE, 2018.